OpenAI's Self-Developed Chip Jalapeño Debuts, Achieving 1.7 Times the Throughput per Watt of GB300
Source：
www.ithome.com
Performance data for the previously reported OpenAI in-house inference chip Jalapeño was publicly disclosed for the first time today. In tests across the GPT-OSS 120B, DeepSeek R1 670B, and Kimi K2.5 1T models, Jalapeño's AI throughput per watt was 1.5 to 1.9 times higher than that of the GB200 and GB300, reaching 1.7 times that of the GB300 specifically under the DeepSeek R1 workload. Its end-to-end latency is approximately 28% to 59% of that of the compared systems.
